SN2410 selection lessons from a documented 25GbE Spark cluster

Overview

NVIDIA’s 2020 Spark 3.0/RAPIDS reference deployment provides a concrete SN2410 application: three physical nodes use 25GbE DAC links, while a separate management path carries administrative traffic. The guide describes preparing both the hosts and switch for RoCE. Its illustrated bill of materials names MSN2410-CB2F, so the example establishes a family use case rather than an exact deployment of MSN2410-CB2RC. For the offered CB2RC configuration, the current hardware ordering table is the better source for the purchase specification: 48 SFP28 ports, eight QSFP28 uplinks, Cumulus Linux, two AC supplies and connector-to-power airflow. Keep that configuration distinct from the P2C/Onyx variant in the historical illustration. The old BOM’s port-count text is inconsistent with the hardware manual and should not be copied into a product page. The useful selection lesson is to start with the complete path. Check the NIC protocol and line rate, cable ends, switch port mode and software configuration together. When adding 100GbE uplinks, calculate the desired rack-level oversubscription from the active server links instead of treating cage count as throughput. This historical guide is a worked deployment example, not a current software baseline or a performance promise. SN2410-CB2RC is now marked EOL by NVIDIA; an existing installation or spare strategy should include software entitlement, tested interoperability, airflow and support planning. Scope: SN2410 family application; reference BOM uses CB2F, not candidate CB2RC